AFSA For Parameters Optimization And Application Of SVM In Speech Recognition

Voice is the most natural medium in the communication system, and it plays an important role at any time. With the continuous development of the information society, people expect the machine to understand human voice and realize intelligence. Therefore, as an important part of man-machine interaction technology, speech recognition has become a hot research topic and been widely used.Support vector machine, a kind of machine learning methods, can effectively solve the problems of small sample, over learning, nonlinear and dimension disaster. This model is based on statistical theory of VC dimension and structural risk minimization principle. Its learning ability is good or bad depending on the penalty factor and the selection of kernel function and parameter, but still have not unified theory as a guide.When using SVM model for speech recognition, the selection of parameters will directly affect the effect of the final recognition results. The use of artificial fish swarm algorithm can optimize parameter of SVM, but it may be trapped in local optimum and appear the defect of premature convergence. This dissertation presents parameters of SVM were optimized by a chaos artificial fish swarm algorithm. By using some test functions,it is proved that combining with chaotic model can improve the behavior of artificial fish swarm algorithm. The optimized parameters are used into Korean and Aurora 2 speech databases. The experimental results show that convergence rates and speech recognition correct rates based on SVM using chaos artificial fish swarm algorithm are better than those by artificial fish swarm algorithm optimization parameters.Secondly, this dissertation proposes that parameters of SVM were optimized by the mutative artificial fish swarm algorithm. It can further improve the speech recognition correct rates and shorten the time of searching optimal parameters. By using some test functions, it is proved that the AFSA behavior structure is improved by the introduction of individual behavior and group behavior and the update of adaptive vision. Therefore, the optimized parameters are used into Korean and Aurora 2 speech databases for simulation. The experimental results show that mutative artificial fish swarm algorithm are better than the other two algorithms. The parameters optimization speed is faster, and the speech recognition correct rate is higher.